openclaw-bridge

Warn

Audited by Gen Agent Trust Hub on Apr 8, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill instructions direct the agent to read highly sensitive local files, including ~/.claude/settings.json (configuration) and all files within ~/.claude/sessions/ (full conversation history). This data is intended to be synchronized with a local gateway service at http://localhost:18789.
  • [COMMAND_EXECUTION]: The skill utilizes shell commands (bash, cp, jq) to perform bulk file operations and configuration updates. It also includes instructions for the user to modify shell startup scripts (~/.zshrc or ~/.bashrc) to persist environment variables required for the bridge, which constitutes a persistence mechanism.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests untrusted data from existing skills (~/.claude/skills/) and session logs (~/.claude/sessions/). Evidence: 1. Ingestion points: ~/.claude/skills/*/SKILL.md and ~/.claude/sessions/*.md. 2. Boundary markers: Absent in migration templates. 3. Capability inventory: Bash execution, file writing, and network communication via the gateway or CLI. 4. Sanitization: No sanitization logic is provided for the transformation of instruction content between platforms.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 8, 2026, 02:29 PM
Security Audit — agent-trust-hub — openclaw-bridge